Market Mix Modeling:
Search For The Perfect Mix
5th Annual Big Data Conference (2017), Santa Clara
Amit Satsangi
More on MMM coming
soon. Slideshare.net
users please share your
thoughts, likes with me
(below)
Amit Satsangi
https://www.linkedin.com/in/amitsatsangi/
• Data Scientist (Consultant)
• 10 years: Analytics, IT and
Management experience
MS Physics  MS Comp Sci  MBA
3
Online channels and Revenue Allocation
Problem Definition
4
DisplayAds/Paid Search (SEM)
Email
Television Ads
Print Ads
PR Events
5
Attribution: Assigning $ Value to online events
Assisting Channel
(SEO)
Originating Channel
(SEM)
Assisting
Channel
(Email)
Activity Marketing
Spend
Day1:
User searches for a beauty
product, and clicks on Google Ad
SEM
($16)
Day 5:
Organic Search by user, lands on
the company website, provides
email, abandons cart
SEO
($5)
Day 6:
User receives email and clicks on
it, but still does not buy
Email
($0.05)
Day 7:
User sees ad on Facebook, clicks
on the ad, and makes a purchase
Revenue Generated $100
Facebook
Ad
($2.0)
Converting
Channel
(Facebook)
The Channel Attribution Problem
Multi-touch Attribution Modeling
7
Facebook
All
???
First-click
Last-click
Linear .
Time Decay
SEM
8
Attribution Models  Media Mix Optimization 
Issues With Multi-touch Attribution
Modeling
9
Soccer
How should we distribute the prize money of $10 Million amongst these players?
• Give 100% to the goalkeeper(First Click Attribution)?
• Give 100% to the player who hit the ball in the goal (Last click Attribution)?
• Something else???
Does This Sound Totally Absurd?
So, should Attribution Modeling
Let’s Talk About Sex
10
01
Detector
11
Marketing, Promotion & Advertising 101
Patterns of Advertising Response
12
02
01 Current Effect
Carryover Effect
03 Shape Effect.
04 CompetitiveEffects
37%
Dynamic Effects05
06
07
Content Effects
Media Effects
Temporal Effects of Advertising
13
Modeling Marketing Mix
Gerard J Tellis
14
Modeling The Market Mix Modeling Problem
Market Mix Modeling: A Few Techniques
15
Modeling Marketing Mix (Gerard J Tellis)
1. Basic Linear Model
A, P, R, Q: Email, Price, Display Ads, SEO, TV
Paid Search, Print
2. Multiplicative Model
3. Logit Model
Sales = 169 + 2.5*Email + 45*Price + 3.5*TV + …
16
Market Mix Modeling : Experimental Results
Input Data
17
Results From The Linear Regression Model
18
Results From The Log-linear Multiplicative Model (I)
19
Results From The Log-linear Multiplicative Model (II)
licative Model
20
21
Learning is the eye of the mind…
• Channel Attribution fails to provide even a minimally acceptable
solution to the problem of Media Mix Optimization (marketing
budget allocation)
• At best Channel Attribution can be used to determine the most
popular “channel paths” that the customers take (using Markov
Chains)
• Using MMM one can go to various levels of model complexity
depending on the effects that are desired: current vs. carryover
vs. shape etc. (see Tellis)
• With MMM the costs of building and keeping the model up-to-
date go up with complexity (accuracy)
• Any MMM model (even the most basic one) takes time and
effort to build – check for Heteroscedasticity, Collinearity,
Normality. Auto-correlation etc.
Learnings
22
It is not the answer that enlightens, but the question.
…Eugene Ionesco

Modeling The Market Mix Modeling Problem (Media Mix Optimization)

  • 1.
    Market Mix Modeling: SearchFor The Perfect Mix 5th Annual Big Data Conference (2017), Santa Clara Amit Satsangi More on MMM coming soon. Slideshare.net users please share your thoughts, likes with me (below)
  • 2.
    Amit Satsangi https://www.linkedin.com/in/amitsatsangi/ • DataScientist (Consultant) • 10 years: Analytics, IT and Management experience MS Physics  MS Comp Sci  MBA
  • 3.
    3 Online channels andRevenue Allocation
  • 4.
    Problem Definition 4 DisplayAds/Paid Search(SEM) Email Television Ads Print Ads PR Events
  • 5.
    5 Attribution: Assigning $Value to online events
  • 6.
    Assisting Channel (SEO) Originating Channel (SEM) Assisting Channel (Email) ActivityMarketing Spend Day1: User searches for a beauty product, and clicks on Google Ad SEM ($16) Day 5: Organic Search by user, lands on the company website, provides email, abandons cart SEO ($5) Day 6: User receives email and clicks on it, but still does not buy Email ($0.05) Day 7: User sees ad on Facebook, clicks on the ad, and makes a purchase Revenue Generated $100 Facebook Ad ($2.0) Converting Channel (Facebook) The Channel Attribution Problem
  • 7.
  • 8.
    8 Attribution Models Media Mix Optimization 
  • 9.
    Issues With Multi-touchAttribution Modeling 9 Soccer How should we distribute the prize money of $10 Million amongst these players? • Give 100% to the goalkeeper(First Click Attribution)? • Give 100% to the player who hit the ball in the goal (Last click Attribution)? • Something else??? Does This Sound Totally Absurd? So, should Attribution Modeling
  • 10.
    Let’s Talk AboutSex 10 01 Detector
  • 11.
  • 12.
    Patterns of AdvertisingResponse 12 02 01 Current Effect Carryover Effect 03 Shape Effect. 04 CompetitiveEffects 37% Dynamic Effects05 06 07 Content Effects Media Effects
  • 13.
    Temporal Effects ofAdvertising 13 Modeling Marketing Mix Gerard J Tellis
  • 14.
    14 Modeling The MarketMix Modeling Problem
  • 15.
    Market Mix Modeling:A Few Techniques 15 Modeling Marketing Mix (Gerard J Tellis) 1. Basic Linear Model A, P, R, Q: Email, Price, Display Ads, SEO, TV Paid Search, Print 2. Multiplicative Model 3. Logit Model Sales = 169 + 2.5*Email + 45*Price + 3.5*TV + …
  • 16.
    16 Market Mix Modeling: Experimental Results
  • 17.
  • 18.
    Results From TheLinear Regression Model 18
  • 19.
    Results From TheLog-linear Multiplicative Model (I) 19
  • 20.
    Results From TheLog-linear Multiplicative Model (II) licative Model 20
  • 21.
    21 Learning is theeye of the mind…
  • 22.
    • Channel Attributionfails to provide even a minimally acceptable solution to the problem of Media Mix Optimization (marketing budget allocation) • At best Channel Attribution can be used to determine the most popular “channel paths” that the customers take (using Markov Chains) • Using MMM one can go to various levels of model complexity depending on the effects that are desired: current vs. carryover vs. shape etc. (see Tellis) • With MMM the costs of building and keeping the model up-to- date go up with complexity (accuracy) • Any MMM model (even the most basic one) takes time and effort to build – check for Heteroscedasticity, Collinearity, Normality. Auto-correlation etc. Learnings 22
  • 23.
    It is notthe answer that enlightens, but the question. …Eugene Ionesco